数字摄影测量影像数据的GPU并行处理研究

发布时间：2018-03-05 22:21

本文选题：GPU　切入点：CUDA　出处：《兰州交通大学》2013年硕士论文　论文类型：学位论文

【摘要】：进入二十一世纪以来，随着科学技术的不断发展，传感器技术取得了长足的进步。测量获取被测物体影像信息的方法也更加多样化，多传感器测量，不同角度测量，大重叠度测量都是当前遥感技术发展的趋势之一。航空、航天影像的高分辨率、大重叠度，致使影像数据量不断增长，海量数据处理已经成为一种趋势。普通的计算机受限于其存储和计算能力，在海量数据处理面前已经力不从心，使用CPU串行处理的生产方式已经难以满足海量数据的高效生产需求。而GPU的高性能并行计算能力和可编程性则在不断发展，为数字摄影测量中一些算法的并行化提供了很大空间，CUDA的推出更是为GPU在通用计算领域的发展提供了强有力的支撑。本文旨在研究基于GPU的并行计算技术进行影像预处理、影像增强与影像匹配，搭建GPU并行处理平台，将GPU作为核心处理设备，基于CUDA4.0软件开发环境，研究影像数据的GPU并行处理，以此来实现数据的高效处理。本文的研究选取了PixelGrid系统中影像预处理算法、Wallis滤波算法、影像匹配算法作为研究对象，研究以上算法的GPU并行处理设计，着重于算法在GPU中的并行实现方法，并基于GPU进行了性能优化，，以此来提高影像处理算法在摄影测量处理系统中的执行效率。论文完成的主要工作和创新点如下: 1.简单介绍了并行计算平台的相关历史和发展趋势，对数字摄影测量影像数据并行处理的基本模式进行了归纳。系统的论述了GPU的硬件体系架构，CUDA软件编程模型，以及CUDA程序的优化，给出了论文研究所采用的实验平台。 2.提出了一种影像预处理GPU并行算法，将影像的旋转、畸变差改正通过GPU进行并行处理，在重采样操作中，利用GPU进行细粒度并行处理，并根据算法的特点和GPU的体系架构，优化了任务划分与执行配置方案，充分发挥GPU的并行计算优势。 3.提出GPU加速的Wallis影像增强并行算法，算法中加入基于GPU的自适应平滑滤波，利用GPU强大的并行能力，实现了Wallis滤波以及影像自适应平滑的GPU并行处理，在算法中运用了共享存储器对速度进行优化。 4.提出了一种基于GPU的Harris算子影像匹配并行处理方法，在GPU中完成对影像的灰度化、Harris角点提取，重采样、灰度相关匹配，并从线程分配、内存使用、共享存储器(share memory)等方面进行优化，充分的发挥出GPU的巨大并行计算能力。实验结果表明，该方法与CPU串行处理方法相比，其速度得到了明显提升。
[Abstract]:Since 21th century, with the development of science and technology, sensor technology has made great progress. The measurement of large overlap degree is one of the trends in the development of remote sensing technology at present. The high resolution and large overlap degree of aerial and spaceflight images make the amount of image data increase continuously. Mass data processing has become a trend. Ordinary computers, limited by their storage and computing capabilities, have been unable to cope with mass data processing. The production mode of serial processing with CPU is difficult to meet the high efficient production demand of massive data. The high performance parallel computing power and programmability of GPU are developing continuously. It provides a large space for the parallelization of some algorithms in digital photogrammetry and provides a strong support for the development of GPU in the field of general computing. This paper aims to study the parallel computing technology based on GPU for image preprocessing. Image enhancement and image matching, build GPU parallel processing platform, take GPU as the core processing equipment, based on the CUDA4.0 software development environment, study the GPU parallel processing of the image data, so as to realize the efficient processing of the data. In this paper, we select the image preprocessing algorithm and image matching algorithm in PixelGrid system as the research object, study the GPU parallel processing design of the above algorithms, and focus on the parallel implementation of the algorithm in GPU. The performance optimization based on GPU is carried out to improve the efficiency of image processing algorithm in photogrammetric processing system. The main work and innovation of this paper are as follows:. 1. The history and development trend of parallel computing platform are briefly introduced, and the basic mode of parallel processing of digital photogrammetric image data is summarized. The hardware architecture of GPU and the software programming model of CUDA are discussed systematically. As well as the optimization of CUDA program, the experimental platform used in this paper is given. 2. An image preprocessing GPU parallel algorithm is proposed, in which image rotation and aberration correction are processed in parallel by GPU. In resampling operation, fine-grained parallel processing is performed by using GPU. According to the characteristics of the algorithm and the architecture of GPU, a parallel algorithm is proposed. The task partition and execution configuration scheme is optimized to give full play to the advantages of parallel computing of GPU. 3. The Wallis image enhancement parallel algorithm accelerated by GPU is proposed. The adaptive smoothing filter based on GPU is added to the algorithm, and the Wallis filter and the GPU parallel processing of image adaptive smoothing are realized by using the powerful parallel ability of GPU. The shared memory is used to optimize the speed of the algorithm. 4. A parallel processing method of Harris operator image matching based on GPU is proposed. The grayscale image is extracted from GPU, resampling, grayscale correlation matching, thread allocation and memory usage. Share memory is optimized to give full play to the huge parallel computing power of GPU. The experimental results show that the speed of this method is obviously improved compared with the CPU serial processing method.
【学位授予单位】：兰州交通大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：P231.5

【参考文献】